Benchmark: prove multi-repository local review target behavior #558

Type

Quality

Area

Review Quality

Priority

P2 — Medium

Problem

The multi-repository Review Target capability exists specifically to fix a failure mode where isolated single-repository review lacks enough target coverage to reason correctly about a coordinated multi-repository change. Nothing proves the delivered implementation actually detects the defects that motivated it, or that its authorization/isolation boundary holds.

Goal

Using the repository's existing benchmark architecture and fixture conventions, prove the delivered multi-repository Review Target capability detects cross-repository defects that isolated single-repository review misses, and that its membership-is-authorization boundary holds.

Scope

• Cross-repo contract mismatch: repo-a changes a producer contract, repo-b changes the consumer incorrectly; the combined review must identify the mismatch.
• Three-repo coordinated change: repo-a (API/contract), repo-b (orchestration/caller), repo-c (downstream consumer) — include both a correct coordinated implementation and an intentionally inconsistent one.
• Repository-local-correctness-vs-combined-defect case: each individual repository looks plausible in isolation; the defect is only evident when the admitted deltas are reasoned about together.
• Isolation/authorization negative case: an unrelated sibling repository exists locally but is not admitted to the Review Target; it must not influence the review.
• Single-repository regression case: normal N=1 review behavior is unchanged.

Non-Goals

• Redefining capability semantics; this issue proves the implementation, it does not change scope.
• External-context/Add bounded local external contract context for cross-repository compatibility review #133-style fixtures.
• Any new benchmark mechanism; reuse the existing benchmark harness/fixture conventions.

Acceptance Criteria

• Benchmark fixtures exist for all five cases above using the existing benchmark harness/fixture conventions.
• The combined review correctly identifies the cross-repo mismatch and three-repo inconsistency cases, and correctly passes the coordinated-correct and N=1 regression cases.
• The isolation/authorization negative case proves the unadmitted sibling repository never influences the review.
• Benchmark evidence is recorded against the actually-delivered implementation, not a hypothetical contract.

Dependencies

Parent: #555
Depends on: implementation child (fixtures may begin contract-first, alongside documentation; final run closes only after the implementation child is delivered)
Relates: #133

Validation

• Run via the existing benchmark harness; record results per repository convention.

Epic invariant (#555). Membership is authorization: cross-repository reasoning may connect evidence across already-admitted Review Target members, but may never add a repository to the target or expand outside those members. Repository-controlled instructions must never expand membership.